In plain words
Computer-use agents learn by carrying out actions on a computer. But exploring real systems can be costly or risky, and building practice environments by hand takes a lot of work.
WildAWM is a pipeline that creates computer worlds with websites, services and files, then checks tasks that require agents to move between interfaces.
Two Qwen models trained in these worlds scored higher on average than their base versions across eight outside tests of web, service and computer-use tasks.
The research
Computer-use agents can learn by interacting with their environments, but real systems are costly and risky to explore and manually built practice environments are expensive to scale.
Can generated worlds provide cross-interface tasks with outcomes reliable enough to support training and show gains beyond those worlds?
WildAWM builds executable computer worlds and checked cross-interface tasks, giving computer-use agents synthetic places to practice.
The researchers trained two Qwen models in WildAWM’s synthetic worlds, first from successful GLM-5.2 task runs and then through practice on separate tasks scored by the system’s verifiers. They compared each trained model with its corresponding base model using the same evaluation harness.
The eight external tests asked agents to complete web, service and computer-use tasks. Three measured the share passed and five measured mean task scores; the average rose by 8.21 points for Qwen3.5-4B and 10.58 for Qwen3.6-35B-A3B. The authors interpret these gains as transfer beyond the training worlds.
The transfer result covers these eight tests and their reported scoring methods; several tests use only a subset of their full task sets, and Online-Mind2Web excludes websites blocked by access restrictions.
Verifier agreement was measured on 100 sampled rollouts, so it does not establish correctness for every task.
Analysis
Deep
- Original paper authors
- Zhaoyang Wang, Qianhui Wu, Haorui Wang, Simon Yu, Baolin Peng
- Submitted to arXiv
- License
- arXiv non-exclusive